Cloud agent
Runs entirely on a provider’s servers. Convenient to start, but your data, context and keys live with the vendor, and so does the switch: an outage, a change of terms or a suspended account stops the agent, whatever the reason.
Concept
Local-first AI agents
A local-first AI agent runs on your own machine instead of a vendor’s cloud. Here is what that means, how it compares to cloud and self-hosted setups, and why it matters for privacy, cost and control.
Which parts actually run locally?
Local-first is not just about where a program starts. An agent has four moving parts: the control loop that plans and decides, the memory it builds up, the tools it runs, and the language model it thinks with. With Veyllo (VAF) the first three always run on your own hardware. The model is the one place where you choose between local and hosted. That split is the whole difference to a cloud agent, where all four sit with the vendor.
The difference
Cloud vs self-hosted vs local-first
Three ways to run an agent, and where your data lives in each.
Self-hosted agent
Runs on infrastructure you operate, usually a server you manage. Control is on your side, but you run and maintain the server.
Local-first agent
Runs on your own machine, in its own window. The installer sets up a local server and account for you, and your data stays with you. No remote infrastructure to provision or maintain. This is how Veyllo runs by default.
Local-first and self-hosted overlap: both keep control on your side. Local-first specifically means the agent runs where you are, on your device. With Veyllo you can do both, run locally or self-host the harness on your own server.
Local, not single-user
One host, your whole team.
Local-first does not mean one person. Even the desktop app opens to your network with one switch, no separate server install needed. Your machine, a spare box at home, or a company server becomes the host, and everyone on the LAN reaches it over HTTPS.
✓
One switch to share. Flip a toggle in the app (or run
vaf server on) and VAF is reachable across your LAN over HTTPS. A plain desktop install is enough, no separate server needed.✓
A separate account per user. Every person gets their own isolated space with separate memory, tasks and preferences. No one sees anyone else’s.
✓
No subscription per seat. One instance serves the whole team or household. You run it once instead of paying for each user.
Why it matters
What that buys you.
✓
Prompts that never travel. Text that does not leave the machine cannot be logged, retained, used for training or handed over on request. With a local model that covers the entire processing chain, not just storage.
✓
No vendor between you and your work. Nobody can raise prices, retire the model you rely on, change rate limits or suspend an account. A setup that works today still works next year, with the same model weights.
✓
Electricity instead of tokens. A local model has no per-token bill. Long contexts, repeated runs and overnight batches cost the same as an idle machine, which is what makes scheduled and multi-step work affordable.
✓
Works without a connection. On a train or a locked-down network the agent keeps going with a local model, memory and files included.
The one real choice
Local model or hosted model?
Everything else about a local-first agent is settled. This is the decision that is actually yours, and it can be changed later.
✓
Local model. Runs on your own GPU or CPU. Nothing leaves the machine and there is no token bill; speed and quality follow the hardware. Small models handle everyday drafting, summarising and file work well, while hard reasoning wants a strong GPU.
✓
Hosted model. The agent still runs locally, only the model call goes out. That puts frontier models within reach without buying hardware, billed per token, and the prompt plus any attached file reaches the model provider.
✓
Both, switched per task. Most setups end up mixed: the local model for routine and confidential material, a hosted one for the hard cases. Swapping takes one line of configuration, so the choice is never permanent.
How Veyllo does it
What Veyllo puts on top of the model.
Veyllo turns a language model into a working agent locally: a persistent vector memory that remembers across sessions, a full toolset (web, code, files, browser) plus anything you connect over MCP, delegable sub-agents, and sandboxed code execution. Run it on a local model, or point it at the hosted, OpenAI-compatible Veyllo API when you want a cloud backend. It is open-source, AGPL-3.0 with a commercial license available.
FAQ
Questions, answered.
What is a local-first AI agent?
The control loop, the memory and the tools run on your device; only the model may sit elsewhere if you choose a hosted one. A cloud agent puts all of it on the vendor's servers, including the context of every request.
Is a local-first agent the same as self-hosting?
They overlap in who holds control, but not in where the work happens. Self-hosting moves the software to a server you administer; local-first puts it on the machine in front of you, with no server to provision. Veyllo does either, and the self-hosted host can serve a whole team.
Do I need to send my data to the cloud to use Veyllo?
No. Veyllo is local-first: you can run the agent entirely on your own machine, models included. The hosted Veyllo API is optional: it provides the models VAF runs on (text, vision, and speech) if you would rather not run your own.
Can more than one person use a single Veyllo install?
Yes. Turn on network sharing and VAF is reachable across your LAN over HTTPS, even from a plain desktop install. Each person gets their own account with isolated memory and tasks, so one host serves everyone without a per-seat subscription.
Can I use the OpenAI SDK with the Veyllo API?
Yes. The Veyllo API is OpenAI-compatible. Point any OpenAI SDK or HTTP client at api.veyllo.app/v1 and use your Veyllo key. Chat and vision are supported.
Can I build my own AI agents with VAF?
Yes. VAF ships on PyPI: pip install --pre vaf pulls the core runtime, and you can build your own AI agents, tools and solutions on top of the framework, including your own harness. Optional extras such as the server or the memory stack install the same way, and the embedding guide in the VAF docs walks through it step by step. The --pre flag is needed while VAF is in alpha.
What license does Veyllo use?
The Veyllo Agent Framework (VAF) is dual-licensed: AGPL-3.0 for open-source use, plus a commercial license for teams that need different terms.
Which operating systems does Veyllo support?
Veyllo runs on macOS, Windows and Linux, as a desktop app, on a server, or from the terminal.